A migration estimated at 231 person-days, delivered in 13 days. That is the result Salesforce Engineering reports after rolling out agentic workflows with Claude Code across 33 endpoints migrated to a cloud architecture (Salesforce Engineering, June 2026). Agentic coding shifts the software engineer's day-to-day work: sequential assisted coding gives way to a model where the agent executes tasks end to end. This guide covers how Claude Code works, the Claude Agent SDK, the MCP protocol and deployment best practices for teams that are scaling up.
What is agentic coding?
From autocomplete to autonomous agent: how have coding tools evolved?
Development tools have evolved over three generations. The first, autocomplete, offers inline suggestions. The second, contextual chat, answers the question you ask. The third, agentic coding, crosses a different threshold.
With traditional assisted coding, the tool suggests and waits. With Claude Code, the agent reads the project files, plans a sequence of actions and executes each task without letting errors slip through. Task execution replaces passive suggestion. For a development team, this is a change in kind in how tooling is used day to day.
How does a coding agent's reasoning and action loop work?
Agentic coding rests on a loop of reasoning, action, then verification.
- Analysing the project context: code files, git history, dependencies
- Planning the sequence of actions that reaches the goal
- Executing the changes and autonomously fixing the errors it runs into
This cycle repeats without manual intervention, including on the longest workflows.
Where does enterprise adoption of agentic coding stand?
Published data confirms the move from experimentation to production. The 2026 State of AI Agents report finds that 91% of the companies surveyed use AI-assisted coding tools in production, and that 54% of respondents from large organisations say they are very optimistic about agent adoption, against 38% at smaller organisations. The barriers cited come in this order: integration with existing systems for 46% of respondents, data access and quality for 42%, security and compliance for 40%.
This ranking has a practical consequence. The main barrier is not the model, it is the connection to the information system. That is also where the gap between a demo and a use that lasts is decided.
Claude Code in agentic mode: how it works and how it is built
The CLAUDE.md file: the agent's persistent memory and rules
Claude Code reads a CLAUDE.md file at the root of the project to give each session its context. This file holds the lasting instructions: code conventions, business rules, permitted workflows. The agent rereads and interprets it every time it is used.
This mechanism removes the need to restate the context over and over. It is the agent's long-term memory and the first rule of any serious use of Claude Code in production.
Skills, subagents and tools: the building blocks of agentic execution
Claude Code's architecture is organised in three levels.
- Subagents: specialised agents that reason autonomously over a precise scope
- Skills: lightweight, reusable capabilities, for example writing a test or documenting a function
- Tools: connections to external systems, through the Claude Agent SDK or the terminal
This hierarchy breaks complex work down into targeted, traceable operations. Subagents improve consistency on large projects, and each level is validated on a small scope before it is rolled out more widely.
MCP: connecting Claude Code to your information system
The MCP protocol is the open standard designed to connect agents to company data. Claude Code supports it natively, with no extra development.
An MCP server connects a Claude Code instance to a Git repository, a database, team messaging or a business tool. Anthropic handed the protocol over to the Linux Foundation's Agentic AI Foundation in December 2025, alongside Block and OpenAI, with the backing of Google, Microsoft, Amazon Web Services, Cloudflare and Bloomberg. Monthly downloads of the reference libraries have since run into the hundreds of millions.
Multi-agent orchestration: delegating several pieces of work in parallel
The dynamic workflows introduced in June 2026 let Claude Code compose its own sequences and split work across specialised subagents. Each agent reads the content within its scope, executes, then fixes the errors it detects.
You set the goal. Orchestration handles the breakdown, the parallel execution and the verification. This is the mechanism behind the time gap Salesforce reports. It is also the one that needs the most guardrails.
Claude Code, Claude Agent SDK or LangGraph: which tool for which use?
Claude Code: the agentic tool for everyday development
Claude Code runs directly in the terminal, with no extra infrastructure.
Main uses: rewriting codebases, multi-file debugging, writing tests, migrations. For a team that works on code every day, it is the fastest way in.
Claude Agent SDK: industrialising beyond the terminal
The Claude Agent SDK exposes the same infrastructure as Claude Code through an API. It is used to build agentic applications: context management, complex tool-calling loops, business agents called from a product.
The SDK takes agentic coding beyond the individual workstation. It marks the switch from a tool a team uses to an agent the company runs.
LangGraph: stateful orchestration for business processes
LangGraph is a state-graph orchestration framework. Its strengths in the enterprise: persistent memory, human intervention built in by design, step-by-step replay of a run.
For business processes with several actors, approval circuits or case handling, LangGraph structures execution in an auditable way. That auditability is what counts when an agent takes part in a decision that commits the company.
Comparing the three: the criteria that decide
| Criterion | Claude Code | Claude Agent SDK | LangGraph |
|---|---|---|---|
| Where it runs | Developer's terminal | Calling application | Orchestrated service |
| State between two runs | Session and CLAUDE.md file | Managed by the application | Persistent by design |
| Human intervention | At the person's initiative | To be implemented | Built in natively |
| Traceability of a decision | Session history | Depends on the implementation | Step-by-step replay |
| Intended use | Everyday work on code | Agentic product | Business process with real stakes |
The three layers are complementary rather than competing. The rule of use stays the same: validate each building block on a small scope before extending it.
Deploying agentic coding: method and guardrails
Which workflows deserve to be handed to an agent?
Not every process lends itself to it. The best candidates meet four conditions.
- Repetitive work with documented rules
- A need for access to the file system or to code repositories
- Calls to APIs or to terminal commands
- Significant manual time spent on an identified task
The first projects to target are code migrations, automated pull request review and documentation generation. Each new use is tested on a non-critical scope, long enough to validate the rules.
How do you frame security and permissions?
Security and compliance rank third among the barriers cited, at 40% of respondents (2026 State of AI Agents). Three rules cover most of it.
- Minimal permissions on the file system: read-only outside the areas explicitly opened up
- Secrets held in environment variables, never written into the CLAUDE.md file
- Control over which MCP servers can be reached, depending on the execution context
These rules are checked at the first deployment, not after the first incident.
What teams that make it last do
The feedback published by Salesforce and Mercado Libre converges on five practices.
- Write the rules in the repository rather than repeating them in each session
- Measure before rolling out, on a comparable scope
- Keep human approval on actions that affect production
- Document every failure in the rules file, so that it does not happen again
- Treat tooling as a team matter, not a personal preference
These practices hold whatever the size of the organisation. They have one thing in common: they turn an individual gain into a team gain.
Three documented enterprise uses
-
Codebase migration. Rakuten reports a refactor carried out over seven consecutive hours in vLLM, an open source library of 12.5 million lines, in a single autonomous run (Rakuten, 2025).
-
Multi-file debugging. The agent reads the dependencies between files and fixes each error at its source rather than where it shows up.
-
Tests in continuous integration. Automatic generation and execution at every commit, with each error documented alongside a proposed fix.
What this shift changes for leadership
The real question is not the tool, it is actual usage
An article on agentic coding speaks first to technical teams. But the decision to extend these uses, pay for the seats and choose between platforms is made higher up. And at that level, the question is no longer "which tool" but "what is actually used, by whom, for what return".
That is precisely what vendor dashboards do not tell you. They cover the scope of their own tool. What escapes them, by design, is what teams do elsewhere, and how the same use compares from one tool to another.
Measure, then decide what to industrialise
We analyse actual AI usage across your estate, from the logs and exports your platforms already produce, without installing anything on workstations. Interviews and data, never one in place of the other. We analyse usage, never people.
The analysis produces three deliverables. The map: what teams really do, including the uses nobody was looking for. The account: what each use case costs and what it returns, paid licences set against observed usage. The course: what to industrialise first, what to cut, who to train on what.
On the estates we analyse, 20 to 40% of paid AI licences go unused. That is a finding from our engagements, not a market average.
What tooling changes after measurement
Measuring is not enough. What moves usage is three tooling levers: integrations that bring the agent to where the work happens, shared assets that turn an individual discovery into a team capability, and shared agents that carry the recurring tasks. From the least costly integration to the one that commits the most.
Targeted and measured in your organisation: 30 to 50% lower cost on the targeted uses, errors divided by 2 to 4. These are engagement targets, not results guaranteed in advance.
Idun Group has been working on these topics since late 2023. At LVMH, for all the perfume and cosmetics brands, an Idun engagement on their generative AI platform in production. At SNCF Immobilier, the audit run in its social-housing subsidiary uncovered 50 critical use cases.
